| paper | d4-conf-aaai-2026-4271 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | medium |
| authors would recognise it | no |
Theorem 8 is a named computational result, but it concerns finite paths of unilateral, single-agent CIS deviations. In a nonatomic population, such a deviation has no aggregate effect; the proposed positive-mass cohort moves change the unit of deviation and the path semantics, so they pose a new problem rather than a continuous version of the anchored result.
fails bit b-changes — the continuum limit changes the question
An atomless agent cannot change the aggregate state, so positive-mass cohort moves replace unilateral deviations with coordinated batch moves.
fatal: True
The proposal targets only the ASHG case of Theorem 8's reachability result; it leaves Theorems 3–6 and 9–12 untouched.
The strongest mirror I see is the paper’s reachability question for individually rational CIS outcomes. It fits a regime where coalition formation is genuinely a population process: for example, a large cooperative or project platform where thousands of workers fall into a few role and compatibility types, and members can move between teams only with the required consent. A type is a complete valuation profile over those types; \(\mu_t\) is the fraction of the population of type \(t\). The point is that \(n\) can be very large while the number \(\tau\) of distinct types stays much smaller.
My anchor is Theorem 8, stated in this paper; its detailed proof is deferred to the full version. It says that, for any functions \(f^+:\mathbb N\to\mathbb Q^+\) and \(f^-:\mathbb N\to\mathbb Q^-\) satisfying \(f^+(n)\ge |f^-(n)|\), it is NP-hard to decide whether CIS dynamics in an ASHG, FHG, or MFHG can reach an individually rational partition from a given starting partition—even when valuations are restricted to \(\{f^-(n),f^+(n)\}\). I would mirror just its ASHG case.
Here is the continuous question I would give a prover: Continuous Mass-CIS-IR Reachability. An instance gives a finite type set \(T\), rational masses \(\mu_t>0\) summing to \(1\), valuations \(v_t(s)\in\{-1,+1\}\), and an initial state consisting of finitely many coalitions, each described by its rational mass vector \(x^j\), plus any mass whose members are singletons. A type-\(t\) member of coalition \(j\) has normalized ASHG utility
\[
U_t(j)=\sum_{s\in T} v_t(s)x^j_s,
\]
and a singleton has utility \(0\). A move transfers a positive rational amount \(\delta\) of one type from a source to an existing or newly formed coalition, or to the singleton pool. Interpret the transfer as a fluid limit: every infinitesimal slice must strictly prefer its destination at the state where it moves, and every affected incumbent must consent under CIS. The question is whether some finite sequence of such mass-flow moves reaches a state that is CIS-stable and individually rational, meaning every type present in every coalition has nonnegative utility. The decision variables are the transferred masses and their order; the task is reachability, not welfare maximization.
This keeps the paper’s central ingredients: ASHG composition-based preferences, unilateral incentives represented by arbitrarily small movements of homogeneous mass, CIS consent, and the question of which stable state the dynamics can reach. The author-facing rationale is that a continuous population does not make preferences or outcomes continuous by itself; it makes the agents’ team assignments into masses, with the same dynamics applied to those masses. I expect the problem to remain hard, perhaps Class C: reachability still involves a sequence of incentive and consent conditions, while fractional splitting could change which paths exist. That is a conjecture for the continuous problem, not a consequence of Theorem 8. The mirror covers this one reachability result, not the paper’s broader meta-theorems or its running-time bounds.
The soft spot is substantial: a positive mass transfer is not literally a single-agent deviation. Fluid moves represent many infinitesimal unilateral moves, and that limit can change reachability—especially when a move’s profitability depends on the first few agents entering a coalition. So the paper’s NP-hardness does not transfer automatically. The case for asking this question is that it preserves the authors’ dynamic problem at the scale where a high-multiplicity population is naturally described; whether the answer changes is exactly what the mirror would test.
Theorem 8 is a real computational result, so “there is nothing here to mirror” would be the wrong objection. The problem is that its object is a finite sequence of single-agent deviations. In a nonatomic society, one agent has zero mass: moving that agent changes no coalition’s type-mass vector and therefore changes no aggregate state. The literal continuum version has no dynamics.
The proposed mass-transfer version repairs that by letting a positive mass of identical agents move together. But that changes the move. In the finite game, the first agent’s move changes the coalition before the next agent decides; in a batch move, members can effectively move simultaneously. Their post-move utilities may include one another, and the consent and incentive conditions can depend on whether the move is evaluated before or after the batch. Taking transfers to zero instead leaves a positive mass to move through infinitely many events, so the paper’s question of reaching a stable state after a finite sequence needs a new path and endpoint convention. Those are choices about the dynamics, not just a mass version of its input.
The type model also needs a substantive restriction. In an ASHG, an agent values particular other agents. A type that preserves those preferences may need to record a valuation for every individual; having only a few types therefore requires assumptions such as block-constant, role-based valuations. That is plausible for some team-formation settings, but Theorem 8’s binary valuation restriction does not itself supply it. The proposed typed problem can be worth studying, but it is a newly specified cohort-flow reachability problem; the paper’s theorem does not give a natural finite-to-continuous dictionary for it. The other dynamic results, whose hardness constructions likewise encode sequences and cycles of individual moves, face the same issue.
This is the strongest negative case, and it has a limit: a careful fluid-CIS model with explicit batch semantics could be a sensible computational question in its own right. I cannot honestly rule out every such model or scenario. The case against is that this paper’s central dynamics do not pass to a continuum without choosing a different notion of move and reachability; calling that choice a mirror of Theorem 8 would need more justification than the proponent provides.
The adversarial triple: the proponent anchors on up to three named results; the opponent sees that case and must defeat every anchor; the judge decides which case convinced it. These are the pipeline’s own outputs, generated by tools/triple_run.py — no human edited them. The paper’s own text is not reproduced here beyond the quoted statement above.